Accelarated Optical Character Recognition on Graphics Processing Units

Optical Character Recognition (OCR) is a technique by the help of which the optical characters are identified automatically by a computer. There are many methods for OCR, one of which is neural network that we use. Unfortunately, the long training and testing time of these networks is disturbing, but we healed this problem by mapping our network on graphics card by using Jacket which is the product of Accelereyes group. By so doing, we achieved the speedup of up to twelve factors. Graphics Processing Units (GPUs) have parallel structure containing many cores capable of running thousands of threads in parallel. We train a multi-layer perceptron network using back propagation rule which has a degree of parallelism that is suitable for implementation on new graphics card. We examine the Persian characters that are typed on the new system of Farsi license plates to make a database of characters uses in this system and apply them as train and test data for our network.

Keywords:

Neural network , OCR , GPU , Jacket , CUDA